Qwen AI model by Alibaba offers low-cost alternative to DeepSeek

Amid the major success of China’s DeepSeek, US computer scientists have developed a new reasoning model that has been trained for less than $50 via the open-source technology of Chinese e-commerce giant Alibaba Group Holdings.

Several researchers from Stanford University and the University of Washington have trained the S1 reasoning model on the back of Alibaba's Qwen 2.5- 32b - Instruct model. Their research paper got published last week.

Also present in the team was Chinese-American “AI godmother” Li Feifei, who works at Stanford, South China Morning Post reported.

Quest for 'cheapest' AI model

This comes after China's DeepSeek took the world by surprise by releasing its high-performance and cost-efficient open-source AI model.

The Chinese AI startup left many stunned with the claim that its chatbot was built at a fraction of the cost of those which have been developed by American tech giants.

This raised questions over the billions of dollars which are being spent by the US-based tech companies over the expansion of data centers which they believe are required to unlock the next wave of AI.

What Qwen research says

As per the research papers, the outcome was obtained after the reasoning model was trained with solutions to 1,000 curated questions. The S1 model was also made to undergo the "thinking process" distilled from the Gemini Thinking Experimental model developed by Google.

On January 29 this year, Alibaba unveiled the new version of the Qwen 2.5. It claimed that the AI model outperformed the DeepSeek-V3 and Meta's Llama-3.1-405B.

The Qwen2.5 series was unveiled for the first time in September. The size of the series ranged from 500 million to 72 billion parameters, according to South China Morning Post.

Low cost

To develop the S1 reasoning model, it required just $14 for the cost of running the graphics processing units (GPUs). This is based on the computer noted during the research, which mentions it being trained on 16 Nvidia H100s for 26 minutes.

The chips could be rented for $2 per hour.

Pan Jiayi, a computer scientist at the University of California, Berkeley, said that the key to train reasoning models at such low cost lies in the base model. “Base model quality is the key,” he noted.

The Qwen2.5 series were unveiled by Alibaba’s cloud computing unit in September last year.

FAQs

1. Who are the competitors of the S1 reasoning model?
The major competitor of the S1 AI model is the OpenAI o1 rival.

2. Is the new version of the Qwen 2.5 better than its rivals?
At its launch, Alibaba claimed that its latest Qwen 2.5 outperforms DeepSeek-V3 and Meta's Llama-3.1-405B.